Milestone 2.5.G7 — Phase 2.5 Exit Gate, Observability, and Handoff

Status: Complete
Goal: Track G — Exit / Ops pack-up
Phase: 2.5 — Provider Generalization & Foundation
Estimated effort: 1 week
Completed: 2026-08-25
ADR: ADR-0051

Why This Milestone Exists

Tracks A–F delivered provider generalization, embedder, schema pre-work, and the MCP skeleton. Phase 2.5 is complete when those seams work together as a coherent platform, observability is operable locally (Prometheus/Grafana/Tempo/Loki), and a single verify gate proves the exit criteria. This milestone is the explicit completion gate — analogous to 2.6.2.

Artefacts:

Local LGTM stack — make observability-up with provisioned dashboards (ADR-0051).
make verify-phase25 — maps exit criteria to runnable checks + multi-service e2e.
Thin K8s observability chart — compose parity for kind/minikube (not full app Helm).
Roadmap sign-off — current-state + phase status completed; Phase 3 entry clear.

Phase 2.5 Exit Checklist

Functionality (Tracks A–F):

Anthropic adapter test depth parity with OpenAI (G1.M1)
Registry.For(model) resolves OpenAI + Anthropic without handler changes (G1.M2)
ModelCapability queryable (context window, tools, tokenizer family) (G1.M2)
Self-hosted OpenAI-compatible adapter (G1.M3)
Tokenizer registry with metrics (G2.M1)
Response pipeline zero-stage byte-identical; overhead budget documented (G3.M1–M2)
Embedder ≥2 real backends: TEI + hosted (OpenAI/Cohere) same contract suite (G4; CPU MiniLM remains stub per ADR-0046)
Schema migrations 14–16 (temporal, multi-label, relationship graph) via make db-migrate (G5)
MCP skeleton + Auth gRPC resource-server boundary + ibex.mcp_tool_calls audit (G6.M1 / ADR-0050)

Operations / observability:

make observability-up starts Prometheus, Grafana, Tempo, Loki, OTel Collector, Alertmanager
Grafana provisioned: System Overview, Proxy Critical Path, Auth, Embedder/MCP
Prometheus scrapes proxy, auth, embedder, mcp-memory /metrics
Proxy/auth OTLP traces appear in Tempo when OTEL_EXPORTER_OTLP_ENDPOINT is set
make observability-smoke exits 0 after traffic
Thin Helm chart under infra/helm/observability/ documented for kind/minikube

Verification:

make verify-phase25 exits 0 (CI job verify-phase25)
Multi-service e2e: auth + proxy + embedder + mcp-memory (make e2e-phase25)
MCP stub conformance evidence (make mcp-conformance)
e2e-wave2b-token-fks asserts current Postgres migration head (16)
Phase 1 security-integration + Phase 2 smokes still green (CI)

Documentation:

Exit criteria on phase index reconciled to shipped reality
current-state reflects Phase 2.5 complete / Phase 3 next
MONITORING.md / OPS_GUIDE / DEVELOPMENT_GUIDE cover observability compose
ADR-0051 accepted; infra README timing updated

Acceptance Criteria

Local LGTM stack + dashboards show live ibex_* series after smoke traffic
make verify-phase25 green and documented
Exit checklist signed off against main
Phase index + overview status → Complete; Phase 3 entry assumes 2.5 exit
K8s observability chart bring-up documented (kind/minikube)

Handoff to Phase 3

Phase 3 builds the memory substrate on Track D/E seams and expands the MCP process (ADR-0050) with real write/search. Org-wide HA observability hardening remains later; local LGTM is the operator path for Phase 3 development.